Home Technology Artificial Intelligence DeepSeek trains AI model on Nvidia Blackwell chip despite USA’s ban

DeepSeek trains AI model on Nvidia Blackwell chip despite USA’s ban

0

On Monday, February 23, 2026, a senior U.S. administration official revealed that DeepSeek—the Hangzhou-based AI startup that has consistently rattled the tech industry—trained its upcoming flagship model using Nvidia’s restricted Blackwell chips.

The disclosure has sent shockwaves through Washington, as the Blackwell architecture is currently under a strict U.S. export ban and is not authorized for shipment to China.


The Inner Mongolia “Phantom” Cluster

U.S. intelligence suggests that DeepSeek’s access to these “crown jewel” semiconductors is not the result of a policy loophole, but a sophisticated evasion of trade controls.

  • The Location: The Blackwell processors are reportedly clustered at a massive data center in Inner Mongolia, an autonomous region in northern China known for its cheap energy and remote infrastructure.
  • Smuggling & Gray Markets: Investigators believe the chips were acquired through a complex “gray market” network involving front companies and recurring deals that bypass official Nvidia distribution channels.
  • Removing Traces: Officials warned that DeepSeek is likely attempting to scrub technical indicators and metadata from its system that would reveal the use of American hardware in its training logs.

Strategic Context: Efficiency vs. Brute Force

What makes this development particularly alarming to U.S. policymakers is that DeepSeek is already known for doing “more with less.”

Model / EventHardware StrategyOutcome
DeepSeek-V3 (2024)Used sanctioned H800 chips.Achieved GPT-4o performance for <$6M.
DeepSeek-R1 (2025)Optimized for lower compute.Proved high-end reasoning doesn’t require H100s.
New Model (2026)Banned Blackwell ClusterSet to release next week; expected to push “frontier” boundaries using banned hardware.

The “Distillation” Allegation

Beyond the hardware theft, the U.S. official echoed recent claims by Anthropic and OpenAI, stating that DeepSeek’s latest model likely relies on distillation—a process where the model is trained on the high-quality reasoning outputs of Western models like Claude and GPT-4.

By combining “stolen” high-end hardware (Blackwell) with “stolen” reasoning data (Distillation), DeepSeek is effectively bypassing the decade-long AI lead that the U.S. export controls were designed to protect.


Political Fallout in Washington

The news has ignited a fierce debate within the Trump administration:

  • The Hawks: Proposing a “total ban” on any AI-related exports and potentially blacklisting cloud providers that allow Chinese entities to rent GPU power remotely.
  • The Pragmatists: White House AI Czar David Sacks and Nvidia CEO Jensen Huang have previously argued that shipping slightly “scaled-down” chips might actually discourage Chinese firms from building their own domestic alternatives (like Huawei’s Ascend 920B).

“This shows why exporting any AI chips to China is so dangerous. Given China’s leading AI companies are brazenly violating U.S. export controls, we obviously cannot expect that they will comply with U.S. conditions.” — Chris McGuire, former National Security Council official.

NO COMMENTS

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Exit mobile version